Case Study of Image Feature Extraction using Conventional and Deep Learning Methods: Autoencoder and U-Net

This paper presents a brief review of traditional and deep learning techniques and classification of images. The conventional methods increase classification and feature extraction design, that turn improved texture, color and, shape feature performance. Handcrafted features are required by traditional feature extraction methods. In order to extract new features from the training data, deep learning is used. While supervised learning still makes up many current deep learning implementations, unsupervised learning is progressively taking its place in order to deal with data (real-world) without manual human interaction. This paper also outlines the famous unsupervised deep learning algorithms for feature extraction such as Restricted Boltzmann Machines (RBM), Autoencoder and, Deep Belief Network (DBN) first, then shows some open challenges of deep learning. A deep learning algorithms autoencoder and U-net model are used as an example in this study to extract features from images. It explains how, even though u-net outperforms convolutional autoencoder in terms of performance, convolutional autoencoder (AE) generated features generate higher accuracy when applied to classifiers. Authors have been used RNNs or its variants (LSTM, Peephole connection, GRU) with CNN as a hybrid model to achieve remarkable effects in image captioning. Following the success of these remarkable outcomes, the attention mechanism developed. A detailed and updated review discussed and several approaches, evaluation metrics, and applications have been presented in the literature.